Predicting a search engine ranking signal value

1. A method of augmenting a first search engine index that indexes resources from a first collection of resources, the method comprising: maintaining, by one or more computers, a first search engine index that indexes resources from a first collection of resources and associates each of a plurality of resources from the first collection of resources with a respective value of a search engine ranking signal for the resource, wherein the respective value of the search engine ranking signal is text of a particular type corresponding to the search engine ranking signal that characterizes the resource; identifying, by the one or more computers, a first resource, in the first collection of resources, that is (i) indexed in the first search engine index and (ii) for which an actual value of the search engine ranking signal is not available in the first search engine index; processing, by the one or more computers, text from the first resource using a machine learning model, the machine learning model being configured to: process the text to generate text of the particular type that is predicted to characterize the resource; updating, by the one or more computers, the first search engine index by associating the text generated by the machine learning model with the first resource as a predicted value of the search engine ranking signal in the first search engine index; and providing the predicted value for the search engine ranking signal in place of the actual value for the search engine ranking signal to a search engine for use in generating a ranking score for the first resource in response to a received search query.

2. The method of claim 1 , further comprising: identifying a plurality of second resources from a second search engine index that indexes a second collection of resources, each of the second resources being associated in the second search engine index with a respective value of the search engine ranking signal for the second resource; generating training data that includes, for each of the plurality of second resources: text of the second resource, and the respective value of the search engine ranking signal for the second resource; and training the machine learning model on the training data.

3. The method of claim 2 , wherein the second collection of resources is a collection of Internet resources.

4. The method of claim 3 , wherein the first collection of resources is different from the second collection of resources.

5. The method of claim 4 , wherein the first collection of resources is a collection of entity-specific resources.

6. The method of claim 1 , wherein the text generated by the machine learning model includes one or more search queries that users would submit to the search engine to search for the first resource.

7. The method of claim 1 , wherein the machine learning model comprises: an encoder neural network configured to process the text of the first resource to generate an encoded representation of the first resource; and a first decoder neural network configured to generate the text of the particular type that is predicted to characterize the resource using the encoded representation of the first resource.

8. The method of claim 7 , wherein a value of a second search engine ranking signal also used by the search engine in ranking resources in response to received search queries is also not available for the first resource in the first search engine index, and wherein the method further comprises: processing the encoded representation of the first resource using a second decoder neural network configured to generate a predicted value of the second search engine ranking signal using the encoded representation of the first resource.

9. A method comprising: identifying, by one or more computers, a plurality of first resources from a first search engine index that indexes a first collection of resources, each of the first resources being associated in the first search engine index with a respective value of a search engine ranking signal for the first resource, wherein the respective values of the search engine ranking signal are used by a search engine to generate ranking scores for the resources in response to received search queries, wherein each respective value of the search engine ranking signal is text of a particular type corresponding to the search engine ranking signal that characterizes the corresponding resource; generating, by the one or more computers, training data that includes, for each of the plurality of first resources: text of the first resource, and the respective value of the search engine ranking signal for the first resource; training, by the one or more computers, a machine learning model on the training data, wherein the machine learning model is configured to: receive text of a resource, and process the text to generate text of the particular type that is predicted to characterize the resource.

10. The method of claim 9 , further comprising: identifying a second resource, in a second collection of resources, that is indexed in a second search engine index for which a value of a search engine ranking signal is not available, wherein a search engine uses values of the search engine ranking signal in ranking resources in response to received search queries; processing text from the second resource using the trained machine learning model, the trained machine learning model being configured to: process the text to generate second text of the particular type that is predicted to characterize the second resource; and updating, by the one or more computers, the first search engine index by associating the second text generated by the machine learning model with the second resource as a predicted value of the search engine ranking signal in the first search engine index; and providing the predicted value for the search engine ranking signal in place of the actual value for the search engine ranking signal to a search engine for use in generating a ranking score for the second resource in response to a received search query.

11. The method of claim 10 , wherein the first collection of resources is a collection of Internet resources.

12. The method of claim 11 , wherein the second collection of resources is different from the first collection of resources.

13. The method of claim 12 , wherein the second collection of resources is a collection of entity-specific resources.

14. The method of claim 10 , wherein the predicted value of the search engine ranking signal for the second resource includes one or more search queries that users would submit to the search engine to search for the second resource.

15. The method of claim 10 , wherein the machine learning model comprises: an encoder neural network configured to process the text of the second resource to generate an encoded representation of the second resource; and a second decoder neural network configured to generate the predicted value of the search engine ranking signal using the encoded representation of the second resource.

16. The method of claim 15 , wherein a value of a first search engine ranking signal also used by the search engine in ranking resources in response to received search queries is also not available for the second resource, and wherein the method further comprises: processing the encoded representation of the second resource using a first decoder neural network configured to generate a predicted value of the first search engine ranking signal using the encoded representation of the second resource.

17. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform operations comprising: maintaining a first search engine index that indexes resources from a first collection of resources and associates resources from the first collection of resources with a corresponding value of a search engine ranking signal for the resource, wherein the corresponding value of the search engine ranking signal is text of a particular type corresponding to the search engine ranking signal that characterizes the resource; identifying a first resource, in the first collection of resources, that is (i) indexed in the first search engine index and (ii) for which an actual value of the search engine ranking signal is not available is not available in the first search engine index; processing text from the first resource using a machine learning model, the machine learning model being configured to: process the text to generate text of the particular type that is predicted to characterize the resource; updating the first search engine index by associating the text generated by the machine learning model with the first resource as a predicted value of the search engine ranking signal in the first search engine index; and providing the predicted value for the search engine ranking signal in place of the actual value for the search engine ranking signal to a search engine for use in generating a ranking score for the first resource in response to a received search query.

18. The system of claim 17 , the operations further comprising: identifying a plurality of second resources from a second search engine index that indexes a second collection of resources, each of the second resources being associated in the second search engine index with a respective value of the search engine ranking signal for the second resource; generating training data that includes, for each of the plurality of second resources: text of the second resource, and the respective value of the search engine ranking signal for the second resource; and training the machine learning model on the training data.

19. The system of claim 17 , wherein the text generated by the machine learning model includes one or more search queries that users would submit to the search engine to search for the first resource.

20. The system of claim 17 , wherein the machine learning model comprises: an encoder neural network configured to process the text of the first resource to generate an encoded representation of the first resource; and a first decoder neural network configured to generate the text of the particular type that is predicted to characterize the resource using the encoded representation of the first resource.